Advanced settings¶
These parameters are rarely needed. None of them is supported for multiclassification: setting one away from its default for a model with three or more classes raises an error.
lambda_scale_invariant¶
Description¶
Rescale lambda_ by the mean Hessian per object of each iteration instead
of using it as it is.
With large sample weights or exposures the sums of the Hessians in the leaves can be so large
that a value of lambda_ in the usual range has no effect. Rescaled, lambda_ acts as a number
of prior objects, whatever the scale of the weights.
Type
bool
Default value
False
ridge_refit_l2¶
Description¶
The L2 penalty of a fully corrective refit. After the trees are built, all their leaf values are re-solved jointly by regularized IRLS, with the tree structures fixed.
None turns the refit off. Not supported with max_depth above 3.
Type
float
Default value
None (off)
ridge_refit_max_iter¶
Description¶
The maximum number of IRLS iterations of the fully corrective refit. Only used when
ridge_refit_l2 is set.
Type
int
Default value
5
dart_drop_rate¶
Description¶
Turns on DART (Dropout Additive Regression Trees): at each iteration, the trees built so far are dropped with this probability before the new tree is built, and the tree weights are normalized as in DART. The value must be in the range \([0; 1)\).
Type
float
Default value
None (off)
nesterov¶
Description¶
Reserved for Nesterov-accelerated boosting, which is not implemented yet. Setting it to True
raises an error.
Type
bool
Default value
False
prune_refit_full¶
Description¶
Deprecated, and without effect: the deployed model is always trained on all objects. Setting it
to True raises an error. It will be removed in a future release.
Type
bool
Default value
False